长春工业大学学报

实验主要以是否风化的高钾、铅钡硅酸盐玻璃为研究对象，针对风化前后玻璃纹饰、类型以及颜色的关系，使用Kendall tau系数及其P值探究之间的相关性和差异性，其中，纹饰和类型与风化系数分别为-0.003 6和-0.344,颜色与风化系数为0.121。对于分类模型，建立随机森林(Random Forest, RF)及K-means++算法，使用网格寻优对RF超参数进行调节，以此来获得影响样本分类的7个重要指标进行数据降维，对降维的数据使用K-means++算法，结合“肘”部图对类别进行选择，实验以铅钡玻璃为例，得出其可分为5个类。对于预测模型建立基于梯度下降的多因素Logistic回归模型，使用梯度下降思想，对Logistic回归模型的β参数进行迭代寻找，模型的准确率和ROC曲线均为1。利用LIME模型，扰动1%～10%的数据，发现预测模型合理且稳定。为后续研究古代硅酸盐玻璃是否风化、分类和预测提供一定的参考价值。

In this paper, the Kendall tau coefficient and its P value were used to explore the correlation and difference between the glass ornamentation, type and color before and after weathering, in which the ornamentation, type and weathering coefficient were -0.003 6 and -0.344, and the color and weathering coefficient were 0.121. For the classification model, the Random Forest(RF) and K-means++ algorithm are established, and the RF hyperparameters are adjusted by grid optimization, so as to obtain 7 important indicators affecting the classification of the sample, the data dimensionality is reduced, the K-means++ algorithm is used for the dimensionality reduction data, and the categories are selected in combination with the "elbow" part diagram, and the experiment takes lead barium glass as an example, and it is concluded that it can be divided into 5 categories. For the prediction model, a multivariate logistic regression model based on gradient descent is established, and the parameters of the logistic regression model are iteratively searched for by using the gradient descent idea, and the accuracy and ROC curve of the model are 1. Using the LIME model, 1%～10% of the data are disturbed, and the prediction model is found to be reasonable and stable. It provides certain reference value for subsequent research on whether ancient silicate glass is weathered, classified and predicted.